Deep Learning Laboratory

Inside the Transformer

Query (Q): The question or feature constraint this token asks for.
Key (K): The address tags this token advertises to all queries.
Value (V): The contextual payload extracted when Query and Key align.

Interactive mathematical exploration of Self-Attention, Multi-Head weights, and Autoregressive token generation

Input Sentence Preset

Attention Head Specialization

Architecture Blueprint: Transformers process all sequence tokens concurrently using self-attention matrices instead of recurrent loops. Click any pipeline step above or hover matrix cells to trace exact dot-product attention scores.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.